Kimi-K3-Audit-451E is Kimi K3
specialised for security auditing. It keeps 451 of the 896 routed experts in
each layer, which brings the model onto two 512 GB Apple silicon machines at
half the size of the release. On kernel-audit workloads it measures at parity
with the full model. Expert selection was calibrated on kernel source and audit
traces, targeting domain focus rather than compression.
For a single 512 GB machine, use the smaller
Kimi-K3-Audit-240E
build instead.
Security auditing and analysis of source code, with tool use. General
conversation, translation and non-technical writing are out of scope, and a
general-purpose model should be used for them.
corpus
perplexity ratio
accuracy 896E
accuracy 451E
change, pts
XNU kernel C
1.01×
80.8%
80.2%
−0.6
Linux kernel C
1.01×
92.3%
92.0%
−0.3
English prose
6.95×
92.0%
50.6%
−41.4
Across eight code corpora in five languages, pruning changes perplexity by at
most 1.10×, against 1.38× for the 240E build.
Perplexity by code corpus relative to the release
Evaluation
Measured against the original MXFP4 release (896E) on identical tokens. KLD is
KL(896E ‖ 451E) over the full output distribution.
Audit traces
Each of 24 kernel-audit sessions contributes one 1025-token window, rendered
through the model's own chat template so the structure matches deployment.
span
ppl 896E
ppl 451E
mean KLD
median KLD
top-1 agreement
tool call
1.616
1.607
0.024
0.00001
97.0%
think
5.874
6.066
0.068
0.043
86.6%
response
3.728
3.560
0.128
0.008
90.8%
other
5.044
5.077
0.207
0.101
79.7%
all
4.232
4.185
0.109
0.023
88.6%
Overall perplexity on audit traces is slightly below the release. Tool-call
spans diverge least of any span measured. 81% of tool-call tokens fall below
0.01 divergence, against 24% of thinking tokens.
This model requires Kimi K3 support from
mlx-lm PR #1626, which has not
yet been merged. Until it lands in an mlx-lm release, install from the PR
branch on both machines.
The model does not fit on one 512 GB machine. Serve it across two with tensor
parallelism, using an MLX distributed
hostfile for the pair. A fast interconnect matters, Thunderbolt RDMA works
well.
Prefill memory scales with chunk × context and is separate from the weights,
so very long prompts may need a smaller --prefill-step-size than the default
2048.
Sampling follows the base model, temperature 1.0 and top-p 0.95.
Limitations
Evaluation is teacher-forced throughout and measures next-token prediction on
reference text. Generation-time behaviour, including loop rate and stop-token
reliability, was not evaluated.
Standard downstream benchmarks were not evaluated.
